What AI Rendering Can and Can't Do for Architects — An Honest 2026 Guide

Most marketing for AI architecture tools describes what the demo did once. This guide describes what the tools do repeatably, based on running them against real projects and keeping the results — including the features we built, tested, and removed because they didn't hold up. If you're deciding where AI fits in your practice, this is the split that matters.

What it does reliably

Today's image models are excellent at changing the appearance of a space while keeping its structure. Everything in this list works consistently enough to put in front of a client:

• Restyling an existing render, elevation, massing study, or photo — new materials, new palette, new level of finish — with the geometry held in place.
• Lighting and atmosphere: day-to-dusk conversion of an exterior photo, relighting an interior, replacing a flat sky, changing the season on a landscape.
• Interior styling: the same room rendered in a different design language, with furnishing direction and lighting guidance.
• Virtual staging: furnishing an empty room photo with furniture and decor (100 credits a photo) for listings and client previews.
• Concept generation from text alone, for early mood and direction work before anything is drawn.
• Finishing work: enhancing materials and lighting on a draft render, upscaling to 4K, regenerating a masked region, erasing an object.
• Colouring a plain floor plan into a furnished presentation plan with the wall layout preserved.

What it does not do — on any current provider

These are not gaps in one product. They are properties of the current generation of image models, and anyone telling you otherwise is describing a demo.

• Removal. Asking a model to empty a furnished room reliably produces a differently-furnished room. We built a "vacant" preset, strengthened the prompt twice, and removed the feature. Use a dedicated eraser on specific objects instead.
• Repurposing. A bedroom does not become a home office on request; the model keeps the room's function even when told the new one explicitly.
• Non-photoreal media. Watercolour, ink, marker, and blueprint styles come back as photographs with a filter on them. For presentation drawings, use a drawing tool.
• Geometry you didn't lock. Without a structural constraint, the same sketch can come back as a different building between runs. Use a structure lock, and don't rely on prompt wording to hold proportions.
• Measurable accuracy from a photo. A 3D model generated from photographs is concept-grade: convincing massing and texture, a single fused mesh, not survey data. It's for visualising and sharing, not for documentation.
• Coherent panoramas. We measured adjacent views from image models disagreeing with each other by a wide margin; a convincing 360 has to come from a real 3D scene, not a stitched set of generated images.

The practical split

Our recommendation, and how Archify is built: do the geometry in a tool that owns geometry, and do the appearance in a tool that owns appearance. Draw walls, openings, levels, and roofs in Design Studio — a real model with rooms, schedules, elevations, and DXF/IFC export, free on every account — and send views of that model to the renderer for materials, light, and mood. The renderer never gets to decide where a wall is, because the wall was never its decision to make.

Setting a client's expectations

• Call it what it is: "an AI visualisation of the design direction" — not a render from the construction model.
• Show the input alongside the output. A sketch-and-render pair is more persuasive than the render alone, and it makes the AI's contribution legible.
• Keep anything dimensioned out of the AI image. Put dimensions on the drawing set the image came from.
• If a render is used in a listing, follow your local MLS rules on labelling virtually staged or AI-altered photos. That disclosure is the agent's responsibility; no tool adds it for you.